|
|
|
|
|
|
|
advantageous, linked genes through a small population. Moreover the study suggests general relations between the number of genes, crossover probabilities, and the rate of adaptation under epistasis. Equally important, the study makes it clear that quite complex ("molecular") definitions of phenotype can be simulated without losing relevance, up to and including suggestions for experiments in vivo and in vitro. (At least two subsequent detailed studies of biological cells were directly encouraged by this experience, R. Weinberg's Computer Simulation of a Living Cell [1970] and E. D. Goodman's Adaptive Behavior of Simulated Bacterial Cells Subjected to Nutritional Shifts [1972].) |
|
|
|
|
|
|
|
|
The first study based directly upon the theoretical framework was that of Daniel Cavicchio. (J. D. Bagley's The Behavior of Adaptive Systems Which Employ Genetic and Correlation Algorithms [1967] is an earlier study which is a direct precursor of both this study and Frantz's.) The set of structures  is taken to be a broad class of pattern classification devices based on those developed by Bledsoe and Browning (1959) and Uhr (1973). Specifically each device uses a set of detectors to process information presented by the sensors in a 25 by 25 array (cf. section 1.3 and Figures 5 through 7). After an initial "training" period, during which the device  accumulates information about one or more handwritten alphabets, A is tested and scored on its classification of letters from another handwritten alphabet. This score amounts to A's performance rating, its payoff µ(A). The adaptive plan, a version of the  class of reproductive plans (pp. 94-95), generates new detectors (and, in the process, new devices) by using genetic operators which are variations on the operators discussed in sections 6.2 through 6.4. |
|
|
|
|
|
|
|
|
Because of the sophistication of the problem environment, the first objective is to develop some estimate of the task's difficulty vis-à-vis the devices in . Cavicchio does this by testing, in the problem setting, a large number of devices drawn at random from . The observed distribution of performances is Gaussian. For a typical environment (Cavicchio calls it the "difficult task"), the mean score is 17 with a standard deviation of 5. (A perfect score would be 100.) This implies that in 1000 random trials of devices drawn from we can expect the best performance to be about 32. |
|
|
|
|
|
|
|
|
To obtain an idea of the performance of a nonreproductive, but adaptive, plan in the same environment, Cavicchio applied a version of Uhr and Vossler's (1973) "detector evaluation" procedure to the search of . This procedure amounts to identifying inferior detectors and replacing them with "mutated" versions. The best performance observed over a great many runs of 600 trials each was a score of 52; each of the runs ''leveled out" somewhere between the 300th and the 600th trial. This is considerably better than a random search, being 7 standard |
|
|
|
|
|